Real-Time Radar-Based Hand Motion Recognition on FPGA Using a Hybrid Deep Learning Model.

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Radar-based hand motion recognition (HMR) faces challenges like sensor interference, clutter, and small datasets, limiting deep learning (DL) model performance.
• Existing DL models for HMR often require substantial computational resources, hindering real-time deployment on embedded platforms.

Purpose of the Study:

• To develop a novel, real-time HMR framework that overcomes noise and data limitations.
• To enhance the efficiency and accuracy of HMR models for deployment on resource-constrained embedded systems.

Main Methods:

• A pre-processing pipeline involving filtration, squared absolute value computation, and normalization was applied to radar data.
• Time-series radar signals were transformed into binarized images for robust feature representation.
• A hybrid Convolutional Neural Network-Support Vector Machine (CNN-SVM) model was employed for classification, followed by quantization and deployment on FPGA platforms (Xilinx Zynq ZCU102 and KR260).

Main Results:

• The proposed CNN-SVM model achieved a classification accuracy of 98.91%.
• Model parameters were reduced by up to 66% compared to recurrent baselines, with SVM achieving 92.79% test accuracy.
• End-to-end accuracies of 96.13% (ZCU102) and 95.42% (KR260) were achieved on FPGA platforms, with significant reductions in execution time and improvements in throughput compared to PC-based implementations.

Conclusions:

• The developed framework enables accurate and resource-efficient radar-based HMR for real-time embedded applications.
• The binarized image representation and hybrid CNN-SVM architecture enhance model robustness and reduce computational load.
• Successful FPGA deployment confirms the system's viability for diverse embedded environments.